Split Conformal Prediction and Non-Exchangeable Data

This repository contains the code for the paper Split Conformal Prediction and Non-Exchangeable Data.

Setup

• All scripts should be run from the root directory (split-conformal-nonexchangeable/).
• Packages listed in requirements.txt should be installed via your python package manager of choice.
• Python 3.11 is required.

Structure

• src: project source code (pushed to the repository).
• data: raw and processed datasets.
• eval: plots, tables and other output used to evaluate conformal methods.

Usage example

We give below a simple example of how the code in this repository can be used for data generation, conformal prediction and coverage evaluation.

from src.models import ConformalQR, RandomForestQR
from src.utils.data import get_synthetic
from src.utils.eval import empirical_coverage

# Generate data from a nonexchangeable autoregressive process
data = get_synthetic("ar1", N=1500, lags=10, seed=0, phi=0.8)

# Split the data into features and target
X, y = data.drop("target", axis=1), data["target"]

# Partition into training, calibration and test sets
X_train, y_train = X[:500], y[:500]
X_cal, y_cal = X[500:1000], y[500:1000]
X_test, y_test = X[1000:], y[1000:]

# Fit, calibrate and generate prediction intervals for a prescribed coverage level of 90%
cqr = ConformalQR(Model=RandomForestQR, alpha=0.1, seed=0)
cqr.fit(X_train, y_train)
cqr.calibrate(X_cal, y_cal)
y_pred_lower, y_pred_upper = cqr.predict(X_test)

# Evaluate coverage
empirical_coverage(y_test, y_pred_lower, y_pred_upper)

Generate figures and tables

Figures 1 and 4

• Run sh src/paper/synthetic-marginal_coverage.sh to generate the results.
• Run python src/eval/synthetic/plot-marginal_coverage.py --stochastic_process ar1 to plot.
• Run python src/eval/synthetic/plot-marginal_coverage.py --stochastic_process two_state_markov_chain to plot.

Download and process climate data

• Run python src/data/download_temp.py -d 19790101-20221231 to download the data.
• Run python src/data/process_temp.py to generate the climatology datasets.

Figure 2

• Run python src/eval/climatology/split_cp.py to generate the results.
• Run python src/eval/climatology/plot-split_cp-maps.py to plot.

Figure 3

• Run python src/eval/synthetic/plot-rolling_coverage-split_cp-enbpi.py to plot.

Table 1

• Run sh src/paper/climatology_benchmarks.sh to generate the results.
• Run python src/eval/climatology/table-benchmarks.py to create the table.

Figure 5

• Run sh src/paper/theoretical_guarantee.sh to generate the results.
• Run python src/eval/synthetic/plot-theoretical_guarantee.py to plot.

Figure 6

• Run sh src/paper/synthetic-empirical_coverage.sh to generate the results.
• Run python src/eval/synthetic/plot-empirical_coverage.py to plot.

Download and process financial data

• Run python src/data/download_forex.py to download the data.
• Run sh src/paper/forex_datasets-process.sh to generate the datasets.

Figure 7

• Run python src/eval/real/marginal_coverage.py --dataset eurusd to generate the results.
• Run python src/eval/real/marginal_coverage.py --dataset bcousd to generate the results.
• Run python src/eval/real/marginal_coverage.py --dataset spxusd to generate the results.
• Run python src/eval/real/plot-marginal_coverage.py to plot.

Table 2

• Run sh src/paper/real-conditional_coverage.sh to generate the results.
• Run python src/eval/real/table-conditional_coverage.py to create the table.